Before this commit, the DLSCH scheduler would construct the MAC PDU by
reading RLC data into a memory on the stack, and then construct the PDU
with CEs first. There are at least two problems:
- we need to keep track of the exact number of bytes of CEs (cumbersome)
  to calculate the number of MAC SDUs to include
- we needlessly copy data around.

This commit does the following instead:
- write all CEs first (no need of keeping track of this in DLSCH and a
  separate function)
- then read MAC SDUs directly into nFAPI as much as possible or
  necessary, without recopying